Hi Fæ,
Thanks a lot for the ideas !
The ideas you mentioned are awesome, and something I'll definitely look
into !
The second and third ideas mentioned are, I believe, do-able within the
scope of my GSoC. For the first idea to be implemented, as you mentioned
local image analysis would be needed, which we've not planned (But i'll add
it to the "to plan" list :) ). Currently we're planning on downloading the
image and performing the analysis on ToolsLab or a personal computer.
Thank you for the project list ! I was looking for a good dataset to test
things out on and this will be immensely helpful.
Regards
Abdeali JK
On Wed, May 18, 2016 at 5:25 PM, Fæ <faewik(a)gmail.com> wrote:
(Just replying on Commons-l with a non-tech
observation. If more tech
stuff arises I'll add it to Phabricator instead)
This looks like a useful contained project, though a lot to be done in
12 weeks. :-)
I was not familiar with catimages.py. It would be great if using the
module for the preparation or housekeeping of large batch uploads were
easy and not time consuming to try. As Commons grows we are seeing
more donations over 10,000 images and have had a few with over 1m.
Uploads of this size make manual categorization a huge hurdle, so
automatic 'tagging' of image characteristics would be a useful way of
breaking down such a large batch to highlight the more interesting
outliers or mistakes, which can then be prioritized on a backlog for
human review.
For example, in my upload projects I have problems detecting:
* incomplete uploads resulting from server failures. Checksum
comparisons would mean re-downloading files, which would be
unnecessarily bandwidth expensive, but local image analysis would
highlight these.
* uploads that are mostly blank pages in old scanned books. I have a
simple detection process, but it would be neat to have a more common
standard way of doing this.
* distinguishing between scans with diagrams and line
drawings/cartoons, printed old photographs, newsprint and text pages.
It would be great if the testing routines you use during the project
could tackle any of these and be written up as practical case studies.
As well as the Phabricator write-up/tracking of the project, it would
be useful to have an on-wiki Commons or Mediawiki user guide. Perhaps
this can be sketched out as you go along during the project, giving an
insight into what other users or amateur Python programmers might do
to customize or make better use of the module? Having an more easy to
find manual, might avoid others going off on their own tangents using
various off the shelf image modules, when they could just plug in
catimages with a smallish amount of configuration.
P.S. If you would like to test the tool on some large collections with
predictable formats, try looking through <
https://commons.wikimedia.org/wiki/User:Fae/Project list >. The 1/2
million images in the book plates project would be an interesting
sample set.
Thanks,
Fae
On 18 May 2016 at 02:53, Abdeali Kothari <abdealikothari(a)gmail.com> wrote:
Hi,
I'm a student from Chennai, India and my project is going to be related
to
performing image processing on the images on
commons.wikimedia to
automate
categorization. DrTrigon had made the script
catimages.py a few years ago
which was made in the old pywikipedia-bot framework. I'll be working
towards
updating the script to the pywikibot-core
framework, updating it's
dependencies, and using newer techniques when possible.
catimages.py is a script that analyzes an image using various computer
vision algorithms and allots categories to the image on commons. For
example, consider algorithms that detect faces, barcodes, etc. The script
uses these to categorize images to Category:Unidentified People,
Category:Barcode, and so on.
If you have any suggestions and categorizations you think might be
useful to
you, drop in at #gsoc-catimages on freenode or my
talk page[0]. You can
find
out more about me on User:AbdealiJK[1] and about
the project at
T129611[2].
--
faewik(a)gmail.com
https://commons.wikimedia.org/wiki/User:Fae
Personal and confidential, please do not circulate or re-quote.